The Capabilities of Large Language Models in Extracting Unstructured Data From Histopathology Reports

The gold standard for medical data extraction has traditionally been manual; however, this approach is very time consuming, labour intensive, expensive and prone to error. Many approaches to automated data extraction have been explored over the years; however, they have required significant technical knowledge and have not been reliably accurate. Large Language Models (LLMs) have been developed at an astronomical pace and have demonstrated incredible accuracy, and they are continuing to evolve and improve. The significant time and cost savings achieved with LLMs will allow for more efficient research and real-time monitoring of patient outcomes. This review explores how LLMs can be used in medical data collection, including the types of data collected and output given, types of LLMs used, amount of training required, the accuracy, speed, and cost of data extraction, types of errors commonly made, and any security concerns. There are still many challenges to overcome, particularly with reducing hallucinations/fictitious content and other common errors, ensuring patient privacy, handling complex tasks and producing output in clean and usable formats. Health professionals and researchers in the field of dermatology must be well trained and upskilled in the use of these new technologies and should continue to explore and build on what has already been achieved to optimise the use of LLMs in the automated data extraction process.